Signal Decision Support

Pharmacovigilance working toolkit
Tool 1 · Gate decision

Signal validation

Determine whether the information supports formal signal assessment, known-risk handling, or continued surveillance.

Structured clinical judgment

Validation questions

Select the statements supported by the available information.

Suggested disposition
Awaiting assessment

Complete the validation questions to generate a transparent recommendation.

Not evaluated
  • Document the information source and data cut
  • Record the decision and supporting rationale
Tool 2 · Attention, not truth

Signal prioritization

Score five domains against written anchors, 1 to 5 each, for a total out of 25. Prioritization allocates assessment capacity. A signal placed in a lower band has not been judged less likely to be real; it has been judged able to wait longer for the same assessment.

Five domains · total out of 25

Score each domain against its anchors

Anchors are defined at 1, 3 and 5. Scores of 2 and 4 are permitted and denote a position between the adjacent anchors, which the rationale must explain. Score the domains before looking at the total, and write each rationale in the language of the evidence rather than the language of the conclusion.

Clinical impact
1 pointNon-serious, self-limiting, resolves without intervention, no lasting effect on function
3 pointsSerious by regulatory criteria, reversible with recognized treatment, no expected permanent sequelae
5 pointsFatal, life-threatening, or irreversible (permanent organ damage, disability, congenital anomaly) in a substantial proportion of affected patients
—
Strength of evidence
1 pointStatistical or operational alert only, or isolated reports without diagnostic criteria or documented time course
3 pointsCoherent series with criteria and compatible latency in several cases, plus one supporting stream (disproportionality above the screening threshold, class finding, or plausible mechanism)
5 pointsConcordant findings across streams with different biases, including a controlled comparison, together with a positive rechallenge or a demonstrated mechanism
—
Exposure
1 pointSmall, well-characterized population; stable or declining use; single market or restricted distribution
3 pointsTens of thousands exposed; stable or moderately growing use across several markets
5 pointsMillions exposed, mass administration, population-wide use, or a launch trajectory where exposure at the time of action will far exceed exposure at scoring
—
Vulnerability
1 pointNo recognized excess susceptibility in the exposed population; event does not concentrate in any identifiable subgroup
3 pointsA definable subgroup (renal or hepatic impairment, older adults, a common co-medication) carries recognizably greater risk and forms a meaningful share of use
5 pointsHarm falls on those least able to recover or least able to consent to risk: pregnancy and the fetus, neonates and children, people unable to report symptoms, healthy recipients of a preventive intervention
—
Preventability
1 pointNo known predictor, no early marker, no modifiable use factor; the event cannot presently be anticipated
3 pointsPartly preventable: a risk factor or co-medication can be avoided in some patients, or earlier recognition would reduce severity, but no reliable monitoring test exists
5 pointsLargely preventable by a feasible measure (a laboratory test with a known window, a contraindication, a dose cap, a device or packaging correction) that would plausibly convert severe outcomes into avoided or recoverable ones
—

Clinical override

The committee may assign a band different from the one the total indicates, in either direction. An override is a recorded decision, not a correction of the arithmetic, and the reason is mandatory. Repeated overrides of the same kind mean the anchors need revision rather than repeated correction by hand.

Required whenever an override is applied. The reason replaces the band the arithmetic produced and is carried into the decision record.
Company design, not a regulatory requirementThe five domains, the band boundaries and the response targets below are an illustrative company design. GVP Module IX Revision 1 expects that signals be prioritised and that those with important impact on patients or public health receive the highest priority; it prescribes no scoring scheme, no domain set, no weighting and no threshold. Describing an internal scale as mandated is an error of fact.
A band never delays a legal obligationEmerging safety issue notification, which runs to three working days, and expedited case reporting run on their own legal timetables regardless of the band a signal occupies. Prioritization is also not the step at which a validated signal is filed with the EU network: notification of validated signals to EMA and the national competent authorities was deleted with effect from 12 August 2025, and a validated signal is carried through the queue under the company’s own signal management process.
—/ 25
Priority band
Awaiting inputs

Score the five domains to produce a total out of 25, a band, and an illustrative internal target to start assessment.

Not evaluated
Bands: high 20–25, medium 13–19, low 5–12. Urgent is reached by override at any score. All four are company definitions.
DomainScoreRationale
Not scored
Clinical overrideA system with no overrides in a year is not disciplined; it is one in which people have stopped reading the cases.
Clinical impact
Accelerated reviewImmediate actionRoutine monitoringPlanned assessment
Strength of evidence →

The plot shows two of the five domains. The band comes from the total of all five, and the override comes from the committee.

Tool 3 · Screening analytics

Disproportionality calculator

Calculate PRR and ROR with confidence intervals, the expected count, the Information Component with its lower credibility bound, and Pearson chi-square from a 2×2 spontaneous-report table.

Hypothesis generation only

Enter report counts

Use mutually exclusive counts from the same database and data lock.

Interpretation boundaryA disproportionality measure quantifies a reporting pattern within one database. It is not a measure of risk, not an incidence, and not a statement of comparative safety between products. Review case quality, reporting behavior, confounding, masking, and clinical coherence before the number changes any conclusion.
Screening result
Ready to calculate

A disproportionality measure quantifies a reporting pattern within one database. It is not a measure of risk, not an incidence, and not a statement of comparative safety between products.

PRR—
PRR 95% CI—
ROR—
ROR 95% CI—
Expected count Nexp—
χ²—
IC—
IC₀₂₅—

Whether a continuity correction was applied is reported here once the measures are calculated.

Not evaluated
Screening convention, not a regulatory requirementAt least 3 reports, PRR ≥ 2, and chi-square ≥ 4. This cut-point set is a convention adopted by many organizations; no guidance obliges its use. A combination meeting all three is a candidate for review, not a signal.
Count stabilityStability flags appear here once the measures are calculated.
Tool 4 · Exposure-anchored analytics

Observed to expected

Compare an observed count inside a predefined risk window against the count expected from an external background rate, with an exact Poisson interval and an explicit background-rate sensitivity range.

Not a causal statement

Exposure, risk window, and background rate

Person-time = exposed × window ÷ 365.25. Expected = rate ÷ unit denominator × person-time.

Count the cases that meet a predefined graded case definition, not the coded terms retrieved. The risk window is specified in advance from the biological hypothesis, not chosen after inspecting the data.

Background-rate sensitivity

The crude ratio is the first line of a sensitivity analysis, not its conclusion. Enter the published range for the background rate.

Interpretation boundaryAn observed-to-expected ratio compares a count against an expectation derived from external data. It is not a rate ratio from a controlled comparison and it is not a causal statement. The exact Poisson interval carries no allowance for uncertainty in the background rate or in the denominator; the sensitivity rows are where that uncertainty is shown.
Observed to expected
Ready to calculate

Enter exposure, risk window, background rate, and the observed count. Quote the range the assumptions produce, not a single ratio.

Person-time (person-years)—
Expected cases—
Observed / expected—
95% CI (exact Poisson)—
Not evaluated

Background-rate sensitivity

Intervals are shown for the primary row only; the scenario rows vary one assumption with nothing else changed.
ScenarioExpectedO/E
Not calculated
Standing cautionEach dose is credited with a full risk window, which overstates person-time near a data cut. Reporting completeness is unknown, which biases the observed count downward. The background rate comes from a population that may differ from those exposed.
Tool 5 · Temporal pattern

Time to onset

Summarize reported onset intervals as median, interquartile range, and range, and show the shape of the distribution rather than a single number.

Describes reported onsets only

Onset intervals

One value per assessable case, in days from the start of the exposure clock. Separate values with commas, spaces, or new lines.

Define the clock before calculating a single interval: first dose, re-initiation, the most recent dose increase, or the start of an interacting co-medication.
Histogram of reported onset
Bars show the number of reported cases per interval once a series is summarized. Dashed line: median. Dotted line: mean.
Cumulative distribution of reported onset
Step curve showing the cumulative percentage of reported cases with onset by each day, once a series is summarized.
Interpretation boundaryThe distribution describes reported onsets among reported cases. It estimates neither incidence nor risk over time, and a median calculated in a stratum of three cases is a number, not a finding.
Onset summary
Ready to summarize

Quartiles are calculated by the Tukey-halves method, stated because conventions differ.

n—
Median (days)—
Q1–Q3, Tukey halves—
IQR width (days)—
Range (days)—
Mean (days)—
Mean ÷ median—
Quartile method—
Not evaluated
Shape before summaryOnset distributions are bounded below by zero and unbounded above. Report median, interquartile range, and range together, and show the distribution.
Tool 6 · Totality of evidence

Evidence integration

Give each stream the weight its validity for this question earns, and record why. No generic hierarchy is applied and nothing is summed: a randomized trial is not automatically superior for a rare delayed event, and a large database is not superior when its outcome definition does not identify the clinical entity.

Justified weight · no score

The question the sources are weighted against

Four attributes of the question decide which sources can carry weight: frequency, latency, ascertainment, and confounding structure. Fix them before assigning any weight.

Evidence streams

One entry per stream, including those examined and found unhelpful: a stream absent from the table is indistinguishable from one nobody looked at. The weight cannot be left blank, and the justification must reference validity for this question rather than volume. No weight is what a source receives when its design cannot address the question, and it is a different entry from a source that addresses the question and points the other way.

Spontaneous case reports
Can supportExistence of a clinical phenotype, latency, dechallenge and rechallenge, susceptibility features
Cannot supportIncidence, comparative risk, absence of risk
Clinical trials
Can supportComparative incidence of events common enough and early enough to occur in the follow-up period
Cannot supportRare or delayed events; effects in excluded populations
Observational data
Can supportComparative incidence in routine care, absolute rates, subgroup contrasts
Cannot supportValidity beyond that of its outcome algorithm and its confounding control
Literature
Can supportIndependent replication, phenotype description, pooled estimates
Cannot supportIndependence, when the reports are the same cases already held
Mechanism and nonclinical
Can supportBiological plausibility, a candidate pathway, dose dependence in vitro
Cannot supportThat the pathway operates at clinical exposure in patients
Class information
Can supportPrior expectation, susceptibility factors, a phenotype to look for
Cannot supportThat this molecule shares the class property

Triangulation

Confidence rises when sources whose biases run in different directions point to the same clinical association. Four sources sharing a bias have not triangulated anything, and independence must be checked rather than assumed.

Discordance

When two sources disagree, the disagreement is information about how one of them was constructed. Averaging destroys that information. Name the axis that would explain the conflict and test it.

Investigate in order of yield: outcome definition first, risk window second, population and comparator third, exposure and data capture last.
Nothing here is computed into a scoreWeighting is an argument to be written down, not an impression, and a fixed source-weight table would answer the question before it was asked. Two failure modes bracket the discipline: averaging, which lets a null result cancel a positive one, and selection, which privileges the stream matching the conclusion already formed.
Totality of evidence
Awaiting entries

Each stream receives a direction and an explicit weight with its reason. The table is the output; there is no total.

Not evaluated
Source and findingDirection, weight, and reason
No entries yet
  • Fix the question before weighting any source
  • Record why each weight was given, for this question
  • Check triangulation and name any discordance axis
Tool 7 · Structure for argument

Causality structure

The Bradford Hill considerations organize an aggregate causal argument so a reader can see what was examined and what was found. They are not criteria to be met, not a checklist, and not a scale. Aggregate causality asks whether the product causes the event in the exposed population; it is a different question from whether this drug caused this patient’s event, and the answer to one does not aggregate into the answer to the other.

No total is computed

Considerations

For each consideration, record what the evidence shows and the caution that attaches to it. A consideration with nothing to show is a finding and belongs here rather than in the reviewer’s memory.

Temporality and induction
AsksDid exposure precede the event within the interval the hypothesized mechanism predicts? Record the interval from initiation and from the most recent dose change.
Strength and specificity
AsksHow large is the association, and how narrow is the differential for the event? Specificity works in the opposite direction from count: a distinctive event is informative in small numbers.
Consistency
AsksIs the finding replicated across sources that could have differed — countries, reporter types, healthcare systems, study designs, time periods?
Dose–response (biological gradient)
AsksIs there a relation to dose, plasma concentration, duration, or cumulative exposure? The four are not interchangeable, and absence of a gradient is evidence against causality only where the mechanism predicted one.
Dechallenge and rechallenge
AsksDid the event resolve on withdrawal, and did it recur on re-exposure? Absence of rechallenge is not negative rechallenge, and a negative rechallenge does not refute an association.
Plausibility and mechanism
AsksIs there a demonstrated pathway that makes the clinical observation interpretable? Mechanistic novelty is not a reason for doubt; novel targets produce novel toxicities.
Analogy and class
AsksDo other members of the class show the same effect? Shared targets raise prior expectation and tell the reviewer what to look for; absence of a class precedent is not evidence against a molecule-specific finding.

Closing the argument

Complete these in order. The field for evidence opposing the association is not optional: an argument with it empty has either been written backward or has not looked.

Case-level categories are tabulated, never summedWHO-UMC categories and Naranjo scores are ordinal single-case classifications, not probabilities. The distance between unlikely and possible is not a defined quantity, their inputs are dominated by documentation quality, and one shared confounder penalizes every case in a series separately. A mean Naranjo score has no interpretation.
Causal argument
Awaiting entries

Each consideration is recorded with the caution that attaches to it. The output is an argument a reader can audit, not a category.

Not evaluated
No score is computed here, and that is deliberateThe considerations are not independent: plausibility, coherence and analogy often reflect the same underlying mechanistic fact, and scoring each separately counts it three times. They are not commensurable: a positive rechallenge and a published class warning are not units of the same thing. And they carry different weight for different mechanisms, so any fixed weighting is wrong for some drug–event pairs by construction. Summing them would hide the judgment that matters. Hill said plainly that none of the considerations could bring indisputable evidence for or against a causal hypothesis and that none could be required as a sine qua non; assigning points to them is a misuse.
ConsiderationWhat the evidence shows, and the caution
No entries yet
  • Name the consideration the mechanism predicts will be informative
  • Write the opposing evidence before the inference
  • Keep the inference in causal-claim vocabulary
Tool 8 · Structured balance

Benefit–risk integration

Given a newly characterized risk, does the balance remain favorable in the population and conditions of use currently authorized, and where it does not, for whom and under what conditions does it fail? The second half of that question does most of the work, because it separates a proportionate action from a blunt one.

Effects table · subgroup decomposition

Favorable and unfavorable effects

State absolute magnitude per 10,000 treated over a stated period, not a relative risk reduction, and give the benefit estimate the same fields as the risk estimate: source, population, magnitude, time period, uncertainty. The risk is examined by a team that knows every weakness of its own estimate; the benefit is too often lifted from a label and enters without an interval.

Favorable effects

Unfavorable effects

Clinical context and value judgment

Subgroup decomposition

The overall balance can remain favorable while it turns unfavorable for a definable subgroup, a dose, a combination, or a setting of use. Show the arithmetic of the restriction explicitly: how much benefit is forgone by removing the subgroup from treatment, and how much harm goes with it.

The action is a recorded judgment. It does not fall out of the arithmetic above.
A ratio of event counts is an input, not a decisionNumbers needed to treat and to harm are silent on severity, reversibility, and who bears each outcome. The identical ratio describes an unacceptable exchange when the harms are fatal and the prevented outcomes are survivable hospitalizations, and a clearly favorable one when the harms are reversible and the prevented outcomes are deaths. State what the prevented outcome and the caused event are, in clinical terms, before any ratio is placed in front of a committee.
Structured balance
Awaiting entries

The effects table and the decomposition are the output. No single margin number is produced, because no single number carries severity, reversibility, or who bears the outcome.

Not evaluated

Decomposition by susceptibility

NNT, whole population—
NNH, whole population—
NNT, remaining population—
NNH, remaining population—
NNT, susceptible subgroup—
NNH, susceptible subgroup—
Harm removed by the restriction—
Benefit forgone by the restriction—

Enter the population, the outcomes prevented, the harms caused, and the susceptible subgroup with its share of the harms.

An assumption, not a findingThe decomposition assumes that benefit is distributed uniformly across the population and that the stated share of harms falls in the susceptible subgroup. Both are assumptions carried into the arithmetic, and a restriction removing a large share of the harm at the cost of a small share of the benefit is proportionate only if they hold.

Effects recorded

EffectEstimate, evidence, importance, uncertainty, subgroup
No entries yet

Proportionate action ladder

1 · Routine monitoring
2 · Targeted follow-up
3 · Additional study
4 · Labeling and communication
5 · Active risk minimization
6 · Restriction, suspension, or withdrawal
Tool 9 · Traceability

Decision record

Create a reconstruction-ready record linking the evidence, uncertainty, conclusion, action, ownership, and reopening criteria.

Audit-ready structure
Documentation check
Record not generated

Required fields will be checked for completeness.

Incomplete
  • Source and data cut
  • Methods, findings, and limitations
  • Decision, rationale, dissent, and approvals
  • Action owner, due date, and follow-up
Integrated output

Signal assessment summary

A concise decision view assembled from the assessment modules. Review and edit the formal source documents before approval.

Exportable record

Unnamed signal

Complete the assessment modules to populate this view.

0%complete
ValidationNot evaluated
PriorityNot evaluated
Quantitative screenNot evaluated
Observed to expectedNot evaluated
Time to onsetNot evaluated
EvidenceNot evaluated
Causality structureNot evaluated
Benefit–riskNot evaluated
DecisionNot documented

Generated assessment record

Copy, download, or print this structured draft for review.

No assessment has been generated.

The full assessment adds every module input as entered, the scenario-laboratory trail, and the standing caveats, as plain text ready to paste into an assessment report.

Important: This educational decision-support tool does not replace a validated company procedure, qualified medical review, applicable regulations, or regulatory authority guidance. Scores and thresholds are illustrative. Decisions should be based on the totality of evidence, documented uncertainty, patient impact, product context, and appropriate governance.
Hands-on learning laboratory

Real-world signal scenarios

Work through incomplete evidence, make decisions at realistic checkpoints, and compare your reasoning with an expert approach.

3 fictional composite cases from the book
Decision 1 of 4
0 points
Case 1

Torvaxel: delayed drug-induced liver injury

An oral small molecule whose label carries no hepatotoxicity warning generates an unexpected serious hepatic report.

Initial alert

Information available now
    What is the best next decision?

    Expert debrief

      Your decision trail
      No decisions recorded

      The record captures what you decided with the information available at each stage.

      Begin the scenarioSelect an option and submit your first decision.
      Scoring principlePoints reward proportionate action, explicit uncertainty, and appropriate use of evidence—not hindsight or excessive intervention.
      Training-case notice: These are fictional composite scenarios created for education. Product names, patients, counts, and outcomes do not describe an actual safety issue. Real decisions must follow the applicable safety-management plan, validated procedures, governance requirements, and regulatory timelines.
      About · scope and limits

      About this tool

      A working companion to the book, built to be used on a real assessment and explicit about what it is not.

      Last reviewed September 2026

      Signal Management in Pharmacovigilance

      From Detection and Validation to Benefit–Risk Decisions and Regulatory Action

      This decision-support tool is the companion to Signal Management in Pharmacovigilance: From Detection and Validation to Benefit–Risk Decisions and Regulatory Action, by Dr. Hafez Selim, MD, PhD, published by Selim Medical Press. The nine modules follow the book’s decision pathway from validation to the decision record, and the scenario laboratory works three of the book’s cases.

      Every worked example and product is a fictional educational compositeVelantor, Pramtelide, Kalvenzo, Torvaxel, NeuVax-7, Doxeprine, Onkelib, Sertavyn, Quinerol, Renavix, Cardemva, Astoval and Miravent do not exist. Their products, patients, counts, laboratory values and outcomes are educational composites, and none of them describes an actual safety issue or an actual product. Every numerical example is illustrative and is shown with its formula, assumptions and limitations.
      Educational; it replaces nothingThis tool is educational. It does not replace applicable law, current regulatory guidance, medical judgment, or your company’s procedures. The scoring bands, domain sets, thresholds, cut-points and response times in it are illustrative company designs rather than regulatory requirements: GVP Module IX Revision 1 expects that signals be prioritised and prescribes no method, and the screening convention of at least 3 reports with PRR ≥ 2 and chi-square ≥ 4 is adopted by many organizations and obliged by none. Nothing produced here is an automated medical or regulatory determination.
      Terminology the tool holds toAn alert is not a signal. Validation is not confirmation, and neither is causality: confirmation is the step by which a competent authority agrees a validated signal warrants prioritisation and assessment. A disproportionality measure describes reporting within one database and does not measure risk, incidence, or comparative safety at any value it can take.
      Data handling
      Nothing is stored or transmitted

      Everything typed into this page stays in this browser tab. The page itself stores nothing, makes no network request of its own, and closing or reloading it discards every entry. One surface is not covered by that: if you read this page in a browser whose AI assistant supports the Web Model Context API, this page offers that assistant read access to the assessment summary you have entered, and a cloud-backed assistant would take that content off this machine. Copy, download or print an assessment before you leave if it is to be kept.

      No storage · no transmission
      • The page itself sends nothing anywhere
      • An AI browser assistant supporting the Web Model Context API can read the summary entered
      • No cookies, no local storage, no analytics, no account
      • No external scripts, fonts, images or network requests
      • Closing the page discards everything

      Currency

      Last reviewed September 2026. Regulatory content — procedure names, database names, guidance versions and timelines — should be re-checked against current guidance before it is relied on, because it changes between printings.

      Two points that moveThe FDA database is the Adverse Event Monitoring System (AEMS), which replaced the legacy FAERS name on 11 March 2026. Notification of validated signals to EMA and the national competent authorities no longer applies: Article 21(2) of Commission Implementing Regulation (EU) No 520/2012 was deleted by Commission Implementing Regulation (EU) 2025/1466, in force 12 August 2025, and a validated signal is now carried under the marketing authorisation holder’s own signal management process. The emerging safety issue notification, in writing and no later than three working days, is unchanged.
      Citation: Selim H. Signal Management in Pharmacovigilance: From Detection and Validation to Benefit–Risk Decisions and Regulatory Action. Selim Medical Press. Companion decision-support tool, reviewed September 2026.